The newly published draft mentions in [expr.prim.req]/6:
If the substitution of template arguments into a requirement would always result in a substitution failure, the program is ill-formed; no diagnostic required. [ Example:
template<typename T> concept C = requires { new int[-(int)sizeof(T)]; // ill-formed, no diagnostic required };— end example ]
But why can't we guarantee the diagnostic to always fail, rather than skip the diagnostic?
Requirement expressions can do pretty much anything. They can provoke further template substitutions, cascading outwardly through an arbitrary amount of code. And recall that template substitutions constitute a Turning complete language.
So you're asking the compiler to, given a Turing complete program, prove whether there is some input which causes that program to be well-formed. This is just a restatement of the Halting Problem. Just like the Halting Problem, there are simple cases where it's obvious the program halts/doesn't halt. But when you're dealing with a Turing-complete language, it can get arbitrarily complex.
The standard isn't going to force compilers to solve the Halting Problem.